Use std::span for cudf::split/cudf::concatenate calls

[joannamooon]

Summary

cudf is migrating its public APIs from cudf::host_span to std::span (rapidsai/cudf#22588). That migration changes the mangled symbols of several exported libcudf functions, two of which rapidsmpf calls: cudf::split and cudf::concatenate. Because rapidsmpf is consumed prebuilt from nightlies, its compiled librapidsmpf.so references the old host_span symbols and fails to load against the new libcudf with:

ImportError: librapidsmpf.so: undefined symbol:
    cudf::concatenate(cudf::host_span<cudf::table_view const>, rmm::cuda_stream_view, ...)

This PR updates rapidsmpf's call sites to the new signatures so it builds and links against the migrated libcudf.

Changes

All in cpp/src/integrations/cudf/partition.cpp:

• cudf::split — the partition_offsets argument was an explicit cudf::host_span<cudf::size_type const>; changed to std::span<cudf::size_type const> to match the new cudf::split signature.
• cudf::concatenate — no source change; the argument (std::vector<cudf::table_view>) converts implicitly to std::span, so
  recompiling against the new libcudf resolves the symbol.
• Added #include <span>.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[wence-]

Thanks for this. You weren't to know, but we're currently porting the cudf-specific parts of rapidsmpf into the cudf repo (see rapidsai/cudf#22747). If you can retarget this change on top of that once it is merged and close this one that would be great!

[vyasr]

Whoops clicked wrong. I also mentioned this on rapidsai/cudf#22588 (comment) 😄